The application of pattern recognition techniques to manufacturing processes is a rapidly developing technology. Automatic verification of the quality of printed wiring boards (PWB's) using pattern recognition techniques is one potential application in this field. Qualitatively, this problem is finding small, irregular features in an environment of complicated, but larger and well-defined geometric features. In addition to the basic pattern recognition task, stringent performance requirements, both for throughput and accuracy, must be met if actual production usage is expected. The method employed in this study is based on characterizing 5 Ã 5 or 7 Ã 7 element binary patterns derived from the class of PWB's being inspected as good or defective. A database of 80 512 Ã 512 element images of PWB's was constructed and used to determine the number of unique patterns and their rates of occurrence. The major experimental result of this study is that less than 500 of the possible (15/16)224 5 Ã 5 patterns are needed to describe all the border containing patterns in the 80 images. It is also apparent that more patterns would be required if the training database was larger. The small number of patterns needed to represent virtually all of the normal border patterns suggests a two-stage inspection strategy. In the first stage, each border pattern from the PWB being inspected is compared to a previously prepared list.